SPO: Structure Preserving Oversampling for Imbalanced Time Series Classification

Abstract

This paper presents a novel structure preserving oversampling (SPO) technique for classifying imbalanced time series data. SPO generates synthetic minority samples based on multivariate Gaussian distribution by estimating the covariance structure of the minority class and regularizing the unreliable eigen spectrum. By preserving the main covariance structure and intelligently creating protective variances in the trivial eigen feature dimensions, the synthetic samples expand effectively into the void area in the data space without being too closely tied with existing minority-class samples. Extensive experiments based on several public time series datasets demonstrate that our proposed SPO in conjunction with support vector machines can achieve better performances than existing oversampling methods and state-of-the-art methods in time series classification.

Integrated Oversampling for Imbalanced Time Series Classification

“…Compared with our previous preliminary SPO work [26], the current proposed INOS method differs and improves in the following aspects: 1) The oversampling is performed in the signal space with improved efficiency and no risk of artificially introducing variances in the common null space.…”

“…As tabulated in Table 7, using our most populous time series data set, Wafer, our current MATLAB implementation takes an average of 2:1 Â 10 À2 and 1:5 Â 10 À2 second for SPO [26] and our proposed INOS, respectively, to create a synthetic sample of 152 dimensions using an ordinary computer with 2.79-GHz CPU. For Yoga, which has the longest time series length, it took 1:7 Â 10 À1 and 5:0 Â 10 À2 second, respectively, for SPO and our INOS to create a sample of 426 dimensions.…”

A Parsimonious Mixture of Gaussian Trees Model for Oversampling in Imbalanced and Multimodal Time-Series Classification

“…Motivated by the data paucity and multimodality in the minority class (e.g., there may be two distinct failure modes for aircrafts), this paper proposes an oversampling method based on a parsimonious mixture of Gaussian trees model for imbalanced time-series classification. Such a model is shown to: 1) compare excellently with other state-of-the-art methods [1], [3] in terms of classification accuracy; 2) require the estimation of far fewer parameters compared with the existing methods; 3) model multimodal minority classes using mixture distributions; 4) model the dependencies between the points in a time-series explicitly; and 5) have a relatively low computational complexity. Existing techniques for class imbalance can be divided into two categories: those at the algorithm-level [4]- [6] and those at the data-level [1], [7]- [16].…”

“…Such a model is shown to: 1) compare excellently with other state-of-the-art methods [1], [3] in terms of classification accuracy; 2) require the estimation of far fewer parameters compared with the existing methods; 3) model multimodal minority classes using mixture distributions; 4) model the dependencies between the points in a time-series explicitly; and 5) have a relatively low computational complexity. Existing techniques for class imbalance can be divided into two categories: those at the algorithm-level [4]- [6] and those at the data-level [1], [7]- [16]. Algorithm-level approaches correct the class imbalance by incorporating a predefined cost assigned to each class [4].…”

“…This imbalance is one of the key challenges [1], [2] because standard classification methods tend to bias toward the class containing the dominant number of learning samples. However, in many real-world applications, the minority class is often the more informative of the two and, indeed, the class of significant interest for accurate modeling.…”